comfyui-advanced-denoiser

🧹 Advanced Image Denoiser — ComfyUI Custom Node

Edge-preserving image denoising that removes noise without making the image blurry.

The node measures the actual noise level of your image (wavelet-based sigma estimation) and, in smart_auto mode, applies just enough denoising to remove it — no more. An edge-aware detail recovery pass then restores fine texture from the original image along edges only, so flat areas stay clean while detail stays sharp.

The node ships with a custom UI panel (slate/teal theme): a segmented method picker with per-method descriptions, contextual sliders that only show the parameters the selected method uses, and a collapsible advanced section. All values sync to the standard widgets, so saved workflows and API use keep working.

Installation

1. Copy this folder into ComfyUI/custom_nodes/comfyui-advanced-denoiser/
2. Install dependencies: pip install -r requirements.txt
3. Optional, highest quality: pip install bm3d
4. Restart ComfyUI — the node is under image → denoising

Methods

| Method | Best For | Notes | |--------|----------|-------| | 🤖 smart_auto | Most images — start here | Measures noise, auto-tunes NLM. strength 0.5 = exactly the measured level | | 🔍 non_local_means | Photo grain, manual control | Separate luminance / chroma strength | | 🎯 bilateral | Portraits, hard edges | LAB-split, edge-preserving | | 🪞 guided_filter | Fast edge-preserving smoothing | Pure numpy implementation, no extra deps | | 〰️ wavelet | Fine grain | BayesShrink with per-channel auto sigma | | 📐 total_variation | Flat/synthetic/AI images | Chambolle TV — strong but keeps edges | | 🏆 bm3d | Maximum quality (slow) | Needs pip install bm3d; falls back to adaptive NLM | | ⚡ median | Salt-and-pepper artifacts | Impulse noise only |

Quick Start

1. Use smart_auto with strength = 0.1–0.25 (yes, that low — higher over-smooths).
2. Raise detail_recovery (0.3–0.6) to bring texture back — it's edge-aware, so it won't re-add noise.
3. Color noise? Use non_local_means and push chroma_strength up (eyes barely notice chroma smoothing).
4. The noise_report STRING output tells you the measured noise sigma per image — wire it to a text display node to see what the node detected.

Key parameters

• strength — keep LOW (default 0.15). In smart_auto, 0.5 applies exactly the measured noise level; 1.0 doubles it.
• detail_recovery — restores original high-frequency detail weighted by an edge map computed from the denoised image, soft-thresholded against the measured noise floor. Safe to raise.
• luminance_strength / chroma_strength — manual methods only. Luminance smoothing is what causes visible blur; chroma can go 2–3× higher safely.
• blend_original — final mix with the untouched input (0.1–0.2 for a natural look).
• sharpen_mode — optional post-sharpen; luminance_only avoids color fringing.

Requirements

• opencv-python >= 4.8
• scikit-image >= 0.21 (noise estimation, wavelet, TV — strongly recommended)
• numpy >= 1.24
• bm3d (optional — enables the bm3d method)

Testing

A standalone smoke test is included — run it with your ComfyUI venv python:

python test_node.py

It runs every method on a synthetic noisy batch and checks shapes, dtypes, value ranges, and that denoising actually reduces noise.

License

MIT